Scalable and Efficient Approaches to Encrypted Computing via Torus-Based Fully Homomorphic Encryption
Guardado en:
| Publicado en: | ProQuest Dissertations and Theses (2025) |
|---|---|
| Autor principal: | |
| Publicado: |
ProQuest Dissertations & Theses
|
| Materias: | |
| Acceso en línea: | Citation/Abstract Full Text - PDF |
| Etiquetas: |
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| Resumen: | Fully Homomorphic Encryption enables arbitrary computation on encrypted data, offering strong privacy guarantees for cloud-based analytics and machine learning. Conventional schemes like BGV or CKKS offer fast computation, but a slow noise refreshing (bootstrapping) process. The Torus-based Fully Homomorphic (TFHE) encryption scheme offers fast bootstrapping, but mostly constrained to boolean or low-precision arithmetic, limiting its adoption to real-world applications. This dissertation explores three of the main bottlenecks of TFHE: its sequential gate bootstrapping algorithm, the lack of higher precision arithmetic and the lack of packing capabilities. Firstly, we present an optimized gate bootstrapping algorithm implementation designed to minimize the latency of bootstrapping by parallelizing latency-sensitive steps of the External Product evaluation. We combine this technique with the Bootstrapping Key Unrolling approach (Zhou et al., 2018), yielding a speedup of 40% compared to a sequential TFHE implementation. Secondly, we introduce a carry-save-based integer-arithmetic framework within the TFHE context, substituting traditional bit-serial operations with a block-wise approach that mitigates bootstrapping overhead and facilitates efficient addition and multiplication of integers. On multi-core systems, our 16-bit multiplier achieves a 15× speedup compared to state-of-the-art CPU implementations and a 3× improvement over optimized GPU baselines. The technique can also be used for any operations that involve repeated accumulation steps. A 45× speedup is obtained for encrypted vector-dot products and a 35× speedup for small convolution kernels, demonstrating the broad applicability of carry-save techniques in FHE. Lastly, we explore one of the main limitations of the TFHE scheme compared to other FHE schemes, its lack of packing capabilities (i.e., encoding multiple values in a single ciphertext, processed in an SIMD manner). We build on the approach of Liu and Wang (2023), which introduced batching possibilities in the TFHE cryptosystem via homomorphic BFV polynomial evaluation. We perform several algorithmic changes designed to increase the parallelism opportunities in the bootstrapping process. On a single 32-core CPU, our implementation outperforms previous packed TFHE implementations by 4.5× for amortized bootstrapping. We then build an efficient matrix multiplication kernel, and obtain over two orders of magnitude speedup for amortized 8 × 8, 8-bit matrix multiply, compared to regular TFHE parallel implementations. Our approach outperforms previous results obtained from a multi-GPU environment. We also apply our efficient packed TFHE bootstrapping algorithm to Convolutional Neural Network Inference, describing how convolution, activation, and max pooling layers can be executed efficiently, obtaining practical runtimes for fully encrypted use cases. Our results also bridge the TFHE performance gap compared to schemes like BFV or BGV. All these contributions advance practical, high-performance encrypted computation via the TFHE scheme towards real-world applications. |
|---|---|
| ISBN: | 9798293827947 |
| Fuente: | ProQuest Dissertations & Theses Global |